AI-Assisted Chronic Wound Assessment
Computer vision and smart-bandage sensing to track how chronic wounds heal, as part of the SWODDYS project.
Chronic wounds such as venous and arterial ulcers heal slowly, and judging their progress still relies largely on subjective visual inspection.
In the SWODDYS project, together with Uniklinikum Regensburg and PreSens GmbH and funded by the Bayerische Forschungsstiftung, I work on making that assessment objective. On the image side, I localise wounds and segment their tissue types with computer vision. On the sensing side, I analyse pH, O2 and CO2 readings from sensors inside the bandage to follow recovery over time.
Tissue segmentation of one wound by the ground-truth mask and eight benchmark models on the CicatrixVision (CiVi) dataset. The wound photo is hidden for patient privacy.
Highlights
- Built the CicatrixVision (CiVi) dataset: 721 wound images from 66 patients, covering 5 wound types and annotated with 3 tissue classes.
- Benchmarked eight segmentation architectures on CiVi; the best, a UNet with an EfficientNet-B0 backbone, reaches a Dice score of 0.80.
- Used Grad-CAM to interpret what CNNs rely on when classifying venous and arterial ulcers.
Outputs
- Paper: Interpreting Venous and Arterial Ulcer Images Through the Grad-CAM Lens, Diagnostics, 2025.
- Poster: Introducing the CicatrixVision (CiVi) Dataset: First steps into AI-assisted Chronic Wound Assessment, MaD Conference 2025 and AI Day, University of Bayreuth 2025.
- Poster: Exploring Correlation Between Tissue Composition and Healing in Chronic Wounds, AIBE Conference 2024.
Left: CiVi dataset poster (MaD Conference and AI Day Bayreuth, 2025). Right: tissue composition poster (AIBE Conference, 2024). Click a poster to open the PDF.
Student projects I supervised
- Alisha Mund: EDTRS Method for Generating Superpixels for Dermatological Wound Images
- Apu Das: Exploring Active Learning for Wound Tissue Segmentation
- Jule Lang: Evaluating Graph Neural Networks for Node Classification on Superpixel-Based Representations of Wound Images
- Ranganath Pujari: Superpixel-Based Wound Segmentation with Frozen Vision Foundation Models and Graph Neural Network Heads